Re: [PATCH] kthread: add kthread_mod_pending_delayed_work api

[Petr Mladek <pmladek@suse.com>]

On Thu 2021-02-18 22:29:24, Yiwei Zhang wrote:
> > 2. User triggered clean up races with the clean up triggered by
> >   timeout. You try to handle this scenario by this patch.
> Yes, exactly.
> 
> > 3. User does clean up after the clean up has already been done
> >   by the timeout.
> This case is well handled. So only (2) has a potential race.

Just to be sure. Does the user work correctly when the clean up work
is done by the timemout before the user wanted to do the clean up?

> Let me clarify a bit more here. The "clean up" is not the clean up
> when a process tears down, but it's actually a "post-work" to cancel
> out an early "pre-work". The "pre-work" enqueues the delayed "post
> work" for the timeout purpose. That pair of operations can repeatedly
> happen.
> 
> The racing is currently worked around by refcounting the delayed_work
> container, and the later "post-work" will take care of the work
> deallocation.
> 
> I mainly want to reach out to see if we agree that this is a valid API
> to be supported by kthread. Or we extend the existing
> kthread_mod_delayed_work api to take another option to not re-queue if
> the cancel failed.

OK, I could imagine a situation when you want to speed up the delayed
work and avoid this race.

The kthread_worker API has more or less the same semantic as
the workqueue API. It makes it easier to switch between them.

The workqueue API has flush_delayed_work(). It does basically
the same as your code. We should call the function
kthread_worker_flush_delayed_work().

I am personally fine with adding this API. I am going to
comment the original code. Well, there might be a push-back
from other people because there will be no in-tree user.

Best Regards,
Petr